lllyasviel / stable-diffusion-webui-forge

GNU Affero General Public License v3.0
8.69k stars 862 forks source link

[Feature Request]: DUAL GPU #48

Open krigeta opened 9 months ago

krigeta commented 9 months ago

Is there an existing issue for this?

What would your feature do ?

can use a Dual GPU setup.

Proposed workflow

no workflow

Additional information

No response

lllyasviel commented 9 months ago

I added --gpu-device-id : https://github.com/lllyasviel/stable-diffusion-webui-forge/commit/74ff4a9ba95ad1d92d48f7bac5f0de8c0c15e398

krigeta commented 9 months ago

I added --gpu-device-id : 74ff4a9

great, how can I use it so it can able to use two graphics card? --gpu-device-id:0,1 ?

SaturMars commented 9 months ago

I added --gpu-device-id : 74ff4a9

great, how can I use it so it can able to use two graphics card? --gpu-device-id:0,1 ?

--gpu-device-id 0 or --gpu-device-id 1

It seems that the two GPUs can't work together at the same time, but when using multidiffusion to generate huge images, is it worth considering distributing different tiles to multiple GPUs?

krigeta commented 9 months ago

I added --gpu-device-id : 74ff4a9

great, how can I use it so it can able to use two graphics card? --gpu-device-id:0,1 ?

--gpu-device-id 0 or --gpu-device-id 1

It seems that the two GPUs can't work together at the same time, but when using multidiffusion to generate huge images, is it worth considering distributing different tiles to multiple GPUs?

thank you and I don't know about that

yggdrasil75 commented 9 months ago

adding accelerate to inference would be nice, but its also only for linux (accelerate crashes constantly on windows if you can even get it to start at all)

ShadowPlague21 commented 3 months ago

networked distributed inference capabilities when? :-)

jswag245 commented 3 weeks ago

@krigeta, hi, just checking to see if the issue has been fixed/solved to any extent. if it is not possible for two GPU's to work together at the same time, and that answer is satisfactory for you, please close the issue.

yggdrasil75 commented 3 weeks ago

this issue should stay open because it is possible, just not with the current implementation. sdcpp has shown that its possible via cublas, and accelerate would make it possible for pytorch, but nobody has done it for a pytorch based ui such as comfyui and webui, only for cpp and c# implementations so far.

krigeta commented 2 weeks ago

this issue should stay open because it is possible, just not with the current implementation. sdcpp has shown that its possible via cublas, and accelerate would make it possible for pytorch, but nobody has done it for a pytorch based ui such as comfyui and webui, only for cpp and c# implementations so far.

I agree, I was about to comment on it, but thanks @yggdrasil75 for this, I hope someday it will be added, till then people must know that as of now it is not possible so that is why this issue is still open.